Why Diagnose the Source of an Association?
A repeated biomarker can predict an event because it carries useful information about a patient’s future risk. That does not, by itself, mean that changing the biomarker would change the risk. In a prognostic study, a biomarker may be entangled with event risk because clinicians respond to it, because an underlying health state affects both the marker and the event, or because a process leading to the event has already begun to alter the marker. These mechanisms can produce similar-looking associations but call for different interpretations.
The diagnostic task is not to prove a causal mechanism from an observational pattern. It is to ask which mechanisms are plausible given the timing, clinical context, treatment history, and available measurements. More than one mechanism may operate in the same cohort. The earlier material on biomarker feedback, latent health, reverse causation, and time-varying endogeneity establishes these mechanisms; here the emphasis is on applying them to concrete prognostic-study scenarios.
These labels describe pathways, not mutually exclusive patient types. For example, an underlying worsening health state may both lower a biomarker and prompt treatment changes; the evolving state may also precede hospitalization. A useful diagnosis therefore begins with a timeline rather than with a single regression coefficient.
A Framework for Diagnosing a Scenario
Suppose \(B_i(t)\) denotes a biomarker for person \(i\) at time \(t\), \(T_i\) the event time, and \(X_i(t)\) measured covariates. One common prognostic summary relates the event hazard to the marker history. For example, a current-value model can be written as
Here \(\lambda_i(t\mid\cdot)\) is the instantaneous event rate at time \(t\) among people who have not yet had the event, and \(\lambda_0(t)\) is a baseline hazard. The coefficient \(\beta\) describes a conditional association under this model; it is not automatically a causal effect of intervening on the biomarker. A Cox-style interpretation also relies on assumptions such as a correctly specified marker relationship, proportional hazard ratios over time for the included covariates, and censoring that is independent of the event process conditional on the modeled information. Even when these modeling assumptions are reasonable, endogeneity can still prevent a causal interpretation.
For diagnosis, ask in order: what happens before each measurement; what happens between measurement and event; and what clinical process could generate both. The following audit makes those questions concrete.
Record biomarker measurements, treatment decisions, symptoms, and events in their actual temporal order. A measurement taken after an event cannot be treated as a pre-event predictor.
Ask whether clinicians or patients act on a marker result, and whether that action can affect later marker values or event risk. Check treatment and behavior histories rather than assuming they are fixed.
Ask whether an evolving or unmeasured condition could change the marker and independently alter event risk. Consider whether recorded covariates plausibly capture that condition’s severity and timing.
Ask whether symptoms or biological changes preceding the recorded event could already affect the marker. Compare measurements made at different distances from the event, while recognizing that such a pattern does not establish direction of causation.
Separate a predictive conclusion from a causal one. Describe plausible mechanisms and remaining uncertainty instead of choosing a mechanism solely because one association is statistically strong.
Worked Problems
Worked Example: A Marker Triggers Treatment Changes
In a hypothetical prognostic cohort, 80 patients have an elevated biomarker at a scheduled visit. The care team intensifies treatment for 40 of them. During the following six months, 12 of those 40 patients have the event. Among the 40 whose treatment is not intensified, 4 have the event. What does this pattern suggest, and what can it not establish?
First calculate the observed event proportions:
The difference is \(0.300-0.100=0.200\), or 20 percentage points over six months in these hypothetical groups. This is a descriptive contrast, not an estimate of the treatment’s causal effect. Patients selected for intensification may have had worse symptoms or more severe disease, and those features may also raise event risk.
The timing makes feedback plausible: an elevated result precedes a treatment decision, and treatment may affect later biomarker values or event risk. To diagnose this pathway, one would examine the dates and reasons for treatment changes, later marker measurements, and relevant clinical history. The calculation alone does not show that treatment increased risk, that the biomarker caused the event, or that intensification was ineffective. Treatment may have lowered risk relative to what it would otherwise have been, even if the treated group’s observed event proportion is higher.
Worked Example: A Hidden Health State Creates a Crude Association
Consider a hypothetical cohort in which an unrecorded health state is either severe or stable. The biomarker is classified as high or low. Within each health-state group, the event proportions are the same for high and low biomarker values:
| Underlying health state | Biomarker | Patients | Events | Event proportion |
|---|---|---|---|---|
| Severe | High | 80 | 32 | 32/80 = 0.400 |
| Severe | Low | 20 | 8 | 8/20 = 0.400 |
| Stable | High | 20 | 2 | 2/20 = 0.100 |
| Stable | Low | 80 | 8 | 8/80 = 0.100 |
Ignoring health state, the event proportion is \(34/100=0.340\) among patients with a high biomarker and \(16/100=0.160\) among patients with a low biomarker. The crude risk ratio is
Yet within the severe group, the proportions are both \(0.400\), and within the stable group, they are both \(0.100\). In this constructed example, the crude association arises because high biomarker values occur more often in the severe group, which also has higher event risk. The calculation illustrates the mixture identity:
Even when \(P(E\mid B,h)\) is the same across biomarker levels within each health state \(h\), the overall risks can differ if the health-state distribution \(P(h\mid B)\) differs. This is the diagnostic signature of possible latent-health confounding: a common cause could explain the marker–event association. In an actual study, the health state may not be known or measured well enough to verify the within-state equality shown here. The example therefore demonstrates a possibility, not a conclusion about any particular cohort.
Worked Example: A Marker Changes Close to the Event
In a hypothetical study, investigators compare 30 patients who experience hospitalization with 30 event-free patients. For each person, a marker is summarized at a remote visit and again near the corresponding index date. The mean marker values are:
| Group | Remote mean | Near-index mean | Change |
|---|---|---|---|
| Hospitalization group | 5.0 units | 8.2 units | 8.2 − 5.0 = 3.2 units |
| Event-free group | 5.1 units | 5.0 units | 5.0 − 5.1 = −0.1 units |
The difference in mean changes is
The marker’s sharper rise near hospitalization is consistent with a pre-event process that changes the marker. For example, worsening illness may affect both the biomarker and the chance of hospitalization. This is a reason to investigate reverse causation, not proof that hospitalization itself caused an earlier biomarker change. A process leading to the event may begin before the event date, and the observed timing may not identify when that process began.
The comparison also depends on how the index dates and measurement windows were chosen. If measurements occur at different times or under different clinical circumstances for the two groups, the contrast may partly reflect measurement schedules or selection. The finding is therefore best reported as a temporal pattern requiring clinical interpretation, not as a causal estimate.
Interpreting Diagnostics and Avoiding Overreach
The three mechanisms suggest different follow-up questions. Evidence that marker results precede and prompt treatment changes supports a feedback concern. Evidence that marker–event associations diminish when a well-measured health state is accounted for is compatible with confounding by that state, although adjustment can be inadequate if health is measured poorly. A marker change concentrated close to the event supports concern about reverse causation, particularly when clinicians describe a plausible pre-event process. None of these patterns alone proves its proposed explanation.
A useful distinction is between prognostic value and causal effect. A marker can improve prediction precisely because it reflects worsening underlying health, even if changing the marker itself would not change the outcome. Conversely, a marker could be part of a causal pathway but show a weak observed association if treatment responds to it. A prognostic model can be useful for risk prediction without supporting a claim that an intervention on the marker would alter event risk.
When comparing approaches, a crude association summarizes observed risk differences but leaves all three mechanisms unresolved. Adjustment for measured covariates can address confounding only to the extent that relevant causes are measured adequately and modeled appropriately; it does not automatically resolve treatment feedback or a pre-event process. Restricting attention to marker values farther from the event can probe sensitivity to near-event changes, but it changes the prediction target and may discard clinically useful information. A richer longitudinal and event-time analysis can represent repeated measurements and event risk together, but more elaborate modeling does not by itself remove the need for defensible assumptions about unobserved health, treatment, and the timing of biological changes.
Common Mistakes and Practical Limits
- Calling prediction causation. A marker that predicts an event need not be a cause of it. State whether the conclusion concerns risk prediction or the effect of changing the marker.
- Assuming adjustment solves endogeneity. Adjusting for measured baseline variables cannot guarantee control of an evolving, unmeasured health state or treatment decisions that respond to marker history.
- Calling every pre-event association reverse causation. Temporal proximity is a clue, not proof. A common cause may affect both the marker and the later event, and the event itself has not occurred at the time of measurement.
- Ignoring the clinical timeline. If treatment occurs after a marker measurement, it cannot explain that earlier value, but it may affect subsequent values and risk. Keep the ordering explicit.
- Treating diagnostic patterns as formal tests. A large risk contrast, a change after adjustment, or a near-event rise does not identify a mechanism without additional assumptions and evidence.
These diagnoses are limited by the information the study collects. Sparse measurements can miss short-lived changes; treatment histories may omit why a decision was made; and a latent health state may be impossible to measure completely. The appropriate conclusion may therefore be that a mechanism is plausible and that causal interpretation is uncertain, rather than that one explanation has been established.
Check Your Understanding
Use the temporal order and proposed pathways to answer each question. Brief solutions follow the questions.
- A high marker is measured, treatment is intensified a week later, and the marker is lower at the next visit. Which mechanism is directly suggested by this sequence, and what additional uncertainty remains?
- In a hypothetical cohort, event proportions for high and low marker values are 0.30 and 0.15 overall. Within each of two health-severity groups, the proportions are equal across marker levels. What mechanism can explain the crude contrast?
- A biomarker rises only in measurements taken shortly before an event. Give one plausible interpretation and one reason the pattern is not proof of reverse causation.
- Can a biomarker be useful for event-risk prediction while having no established causal effect on the event? Explain.
Solutions. (1) The sequence suggests feedback because the marker result precedes a treatment response that may affect later marker values. It does not establish that treatment caused the later decrease or determine how treatment affected event risk. (2) Latent-health confounding can produce the crude contrast if health severity affects both marker distribution and event risk; the equal within-group proportions show how this can occur in the hypothetical data. (3) A pre-event illness process could alter the biomarker before hospitalization. The rise might instead reflect another common cause, measurement timing, or a different process; temporal proximity alone does not establish the direction of causation. (4) Yes. The marker may reflect underlying risk and improve prediction even if changing the marker itself has no established effect on the event.